Not send, failure during completion, completed - control inconclusive and completed - control inconclusive.
Send out againd and what if the outcome is equal to mine or someebody else's ?
The crunching on the anonymous platform was done with the CUDA application, the other one with the CPU application. The task will sent out to another computer until there is a valid result and depending what "referee application" it will be, one of the actual results will be validated.
If the referee is CUDA, then CUDA will win, if it is the CPU app, then the CPU app will win. The other task will be "marked as invalid".
Same thing happens to me in SETI@home. My CPU result does not agree with a CUDA result and is marked inconclusive. IMHO apples should not be compared with oranges. The CUDA world is different from the CPU world.
Tullio
Same thing happens to me in SETI@home. My CPU result does not agree with a CUDA result and is marked inconclusive. IMHO apples should not be compared with oranges. The CUDA world is different from the CPU world.
Tullio
But there are also results, where CUDA and CPU results are matching
So the question, what conditions cause the mismatch. And as mentioned earlier in this thread, this could be burried in the lower precision of the calculations on the GPU, which gets out of range.
That causes the intersting question if CUDA reults are in general precise enough.
Invalid tasks for SpekAal.
now around 12 pieces and a few on hold that it does not make it.
And there are still a few to go there, more and more.
what can or will be done, we patiently wait
hope for answers soon
So the question, what conditions cause the mismatch. And as mentioned earlier in this thread, this could be burried in the lower precision of the calculations on the GPU, which gets out of range.
That causes the intersting question if CUDA reults are in general precise enough.
This has been debated in the SETI message boards, where CUDA users and results abound more than in Einstein. CUDA/CUDA agreed results might not be valid at all, but this does not matter much in SETI. It might matter more in Einstein or QMC or AQUA or CPDN or LHC. This is why QMC, AQUA, LHC and CPDN have no CUDA version.
Tullio
So the question, what conditions cause the mismatch. And as mentioned earlier in this thread, this could be burried in the lower precision of the calculations on the GPU, which gets out of range.
That causes the intersting question if CUDA reults are in general precise enough.
This has been debated in the SETI message boards, where CUDA users and results abound more than in Einstein. CUDA/CUDA agreed results might not be valid at all, but this does not matter much in SETI. It might matter more in Einstein or QMC or AQUA or CPDN or LHC. This is why QMC, AQUA, LHC and CPDN have no CUDA version.
Tullio
But Primegrid does and I cannot imagine they accept incorrect results.
So it must be the way the E@H WU's are completed at the moment.
Or am I wrong ?
But Primegrid does and I cannot imagine they accept incorrect results.
So it must be the way the E@H WU's are completed at the moment.
Or am I wrong ?
AFAIK Primegrid requires double precision support in its GPU version, which not all graphic cards offer. Since I am not a GPU user I do not know if CUDA supports double precision. I know some ATI cards offer double precision but CUDA does not run on them.
Tullio
RE: What happens with the
)
The crunching on the anonymous platform was done with the CUDA application, the other one with the CPU application. The task will sent out to another computer until there is a valid result and depending what "referee application" it will be, one of the actual results will be validated.
If the referee is CUDA, then CUDA will win, if it is the CPU app, then the CPU app will win. The other task will be "marked as invalid".
Same thing happens to me in
)
Same thing happens to me in SETI@home. My CPU result does not agree with a CUDA result and is marked inconclusive. IMHO apples should not be compared with oranges. The CUDA world is different from the CPU world.
Tullio
RE: Same thing happens to
)
But there are also results, where CUDA and CPU results are matching
http://einsteinathome.org/workunit/91537628
So the question, what conditions cause the mismatch. And as mentioned earlier in this thread, this could be burried in the lower precision of the calculations on the GPU, which gets out of range.
That causes the intersting question if CUDA reults are in general precise enough.
Invalid tasks for SpekAal.
)
Invalid tasks for SpekAal.
now around 12 pieces and a few on hold that it does not make it.
And there are still a few to go there, more and more.
what can or will be done, we patiently wait
hope for answers soon
translated by google
a few examples
http://einsteinathome.org/workunit/91367639
http://einsteinathome.org/workunit/91367577
http://einsteinathome.org/workunit/91327012
http://einsteinathome.org/workunit/91004541
http://einsteinathome.org/workunit/91153230
RE: But there are also
)
This has been debated in the SETI message boards, where CUDA users and results abound more than in Einstein. CUDA/CUDA agreed results might not be valid at all, but this does not matter much in SETI. It might matter more in Einstein or QMC or AQUA or CPDN or LHC. This is why QMC, AQUA, LHC and CPDN have no CUDA version.
Tullio
RE: RE: But there are
)
But Primegrid does and I cannot imagine they accept incorrect results.
So it must be the way the E@H WU's are completed at the moment.
Or am I wrong ?
RE: The cross validation
)
As Mike already pointed out, we're looking into this.
Thanks for your patience,
Oliver
Einstein@Home Project
RE: But Primegrid does and
)
AFAIK Primegrid requires double precision support in its GPU version, which not all graphic cards offer. Since I am not a GPU user I do not know if CUDA supports double precision. I know some ATI cards offer double precision but CUDA does not run on them.
Tullio
RE: AFAIK Primegrid
)
nope they don't..
RE: RE: AFAIK Primegrid
)
I just read it in their Web page. They list 3 nVidia cards which can do double-precsion math:
GTX260-295,400,500
Tullio